Re: pg_upgrade: fail early if a tablespace dir already exists for new cluster version

On Thu, Oct 15, 2020 at 07:35:30PM -0400, Bruce Momjian wrote:
> On Fri, Oct 9, 2020 at 07:42:51PM -0400, Bruce Momjian wrote:
> > On Fri, Oct 9, 2020 at 02:23:10PM -0500, Justin Pryzby wrote:
> > > In my local branch, I had revised this comment to say:
> > >
> > > + * Note, v8.4 has no tablespace_suffix, which is fine so long as the version we
> > > + * being upgraded *to* has a suffix, since it's not allowed to pg_upgrade from
> > > + * a version to the same version if tablespaces are in use.
> >
> > OK, updated patch attached. Also, from your original patch, I didn't
> > need to call canonicalize_path() since we are not comparing paths, and I
> > didn't need to include common/relpath.h. I also renamed a variable.
>
> Patch applied to all supported versions. Thanks for the report, and the

I wonder if pg_upgrade should try to rmdir() the tablespace dirs before
restoring global objects, allowing it to succeed, rather than just "failing
early". That can avoid the need to re-create the new cluster, which I imagine
in some scenarios might be bad enough to require aborting the upgrade.
I propose only for master branch.

This doesn't avoid the need to re-create the new cluster if anything has been
restored into it (eg. if pg_tablespace is populated), it just cleans out any
empty dirs left behind by a previous pg_upgrade which failed after "restoring
global objects" but before copying/linking data into that tablespace.

|rm -fr pgsql12.dat pgsql14.dat tblspc;
|/usr/lib/postgresql/12/bin/initdb -D pgsql12.dat
|./tmp_install/usr/local/pgsql/bin/initdb -D pgsql14.dat
|mkdir tblspc tblspc/PG_14_202010201
|echo "CREATE TABLESPACE one LOCATION '`pwd`/tblspc'" |/usr/lib/postgresql/12/bin/postgres --single postgres -D pgsql12.dat -p 5678 -k /tmp
|./tmp_install/usr/local/pgsql/bin/pg_upgrade -D pgsql14.dat/ -d pgsql12.dat -b /usr/lib/postgresql/12/bin

--
Justin